mistral-nemo-kartoffel-12B

Mahou-1.5-mistral-nemo-12B-lorablated finetuned on various datasets.

Method

ORPO tuned with 8x A100 for 2 epochs.

QLoRA config:

# QLoRA config
bnb_config = BitsAndBytesConfig(
 load_in_4bit=True,
 bnb_4bit_quant_type="nf4",
 bnb_4bit_compute_dtype=torch_dtype,
 bnb_4bit_use_double_quant=True,
)

# LoRA config
peft_config = LoraConfig(
 r=16,
 lora_alpha=32,
 lora_dropout=0.05,
 bias="none",
 task_type="CAUSAL_LM",
 target_modules=['up_proj', 'down_proj', 'gate_proj', 'k_proj', 'q_proj', 'v_proj', 'o_proj']
)

Training config:

orpo_args = ORPOConfig(
 run_name=new_model,
 learning_rate=8e-6,
 lr_scheduler_type="linear",
 max_length=2048,
 max_prompt_length=1024,
 max_completion_length=1024,
 beta=0.1,
 per_device_train_batch_size=4,
 per_device_eval_batch_size=4,
 gradient_accumulation_steps=1,
 optim="paged_adamw_8bit",
 num_train_epochs=2,
 evaluation_strategy="steps",
 eval_steps=0.2,
 logging_steps=1,
 warmup_steps=10,
 max_grad_norm=10,
 report_to="wandb",
 output_dir="./results/",
 bf16=True,
 gradient_checkpointing=True,
)